Material-agnostic machine learning approach enables high relative density in powder bed fusion products

Abstract

This study introduces a method that is applicable across various powder materials to predict process conditions that yield a product with a relative density greater than 98% by laser powder bed fusion. We develop an XGBoost model using a dataset comprising material properties of powder and process conditions, and its output, relative density, undergoes a transformation using a sigmoid function to increase accuracy. We deeply examine the relationships between input features and the target value using Shapley additive explanations. Experimental validation with stainless steel 316 L, AlSi10Mg, and Fe60Co15Ni15Cr10 medium entropy alloy powders verifies the method's reproducibility and transferability. This research contributes to laser powder bed fusion additive manufacturing by offering a universally applicable strategy to optimize process conditions.

Fig. 1. Scatter plot of relative density [%] vs laser energy density [J/mm³] for STS 316 L, Ti-6Al-4V, and AlSi₁₀Mg powders in L-PBF.

Description: The scatter plot highlights the variation in optimal energy density required to achieve high relative density for each material, as well as the differing relative densities achieved with the same energy density when different process parameters are used. The references of the data are given in the supplementary materials. Source data are provided as a Source Data file.

Fig. 2. Feature importance and SHAP analysis of input features without considering interactions.

Description: a Average absolute SHAP scores of input features, highlighting the relative importance of process parameters and material properties of powder. Visualization of the SHAP main scores for b energy density, c scan speed, d laser power, e thermal conductivity, f density, g melting point, h reflectivity, i layer thickness, j hatch spacing, and k specific heat capacity and their effects on the model’s output, relative density. Source data are provided as a Source Data file.

Fig. 3. SHAP analysis of the interactions between input features.

Description: Visualization of the SHAP interaction values for a interaction between laser power and scan speed, b interaction between laser power and thermal conductivity, c interaction between scan speed and thermal conductivity, e interaction between hatch spacing and scan speed, f interaction between density and scan speed, and g interaction between heat capacity per volume and scan speed. d Visualization of the combined SHAP scores, representing the sum of SHAP interaction scores for scan speed and thermal conductivity, and the SHAP main scores for scan speed. h Visualization of the combined SHAP scores, representing the sum of SHAP interaction scores for scan speed and heat capacity per volume, and the SHAP main scores for scan speed. Source data are provided as a Source Data file.

Fig. 4. Optical micrographs of the cut surfaces for the specimens with the highest relative density.

Description: Optical micrographs showing the cut surfaces of the top six specimens with the highest relative density for a STS 316 L, b AlSi10Mg, and c Fe60Co15Ni15Cr10 MEA.